Beyond twitter

Prototyping a data extraction pipeline for bluesky.social and exploration of bluesky user activity for influenza like digital disease detection

Heiner Atze

Digital Epidemiology 2025, Hasselt University

2025-04-10

Outline

The bluesky social network
Data accessiblity via the bluesky API
Extraction and Analysis of ILI related bluesky messages

Introduction

`bluesky`: general aspects

microblogging platform
similar to twitter in user experience
decentralized
open source

Decentralization and Democratization of content algorithms ¹

Decentralized User Identifier (DID)
- immutable, associated with human readable user handle
Personal Data servers (PDS)
DIDs and affiliated contents are portable between PSDs
Users can choose, prioritize and develop feed generators and content labelers

Development of user activity ¹

current estimate: ca. 33 Millions active users
user base expanded in bursts after key events:
- 2022: acquisition of twitter by Elon Musk
- 2024: ban of X in Brazil, presidential election in the US

Literature addressing `bluesky`

Google scholar search : “bluesky” AND “social” since 2022
43 articles
main topics:
- decentralized social network architecture
- user migration from X to bluesky 2024
- network structure and dynamics
no results for
- “bluesky” AND “disease”
- “bluesky” AND “epidemiology”

Exploration of bluesky data

bluesky API

publicly accessible for free
extensive documenation at https://docs.bsky.app/docs/category/http-reference

`searchPosts` API method

API documentation
selected parameters:
- q: search query
- since, until: defining search period
limit: max. 100 posts
deterministic search
allows exhaustive sampling

Post metadata

defined in the SDK documentation
fields (selection):
- uri: unique post identifier
- author: contains did which allows to retrieve user profile
- record: contains the text and time information of the message
  - langs: language(s) detected by the bluesky server
- embedded: any embedded media (images, other posts, etc …)
in contrary to former twitter post metadata, no geoinformation

User information

Feedgens
Labelers
no geo information

`getProfiles` API endpoint

allows to retrieve the author profile information

Project

Outline

bluesky post data for digital disease surveillance

Implementation of a continuous surveillance pipeline

Data extraction

Basal network activity

Keywords:
- travail (work)
- demain (tomorrow)
- voiture (car)
- sommeil (sleep)
post counts aggregated by day

Case data

data downloaded from WHO Flumart
- FluID: centralized epidemiological surveillance data

Results

Post count time series

Raw posts counts

Data analysis starting from 2023-08-01

Keyword posts vs. ILI incidence

	ILI posts	Control posts	ILI incidence
ILI posts	1.000	0.863	0.773
Control posts	0.863	1.000	0.541
ILI incidence	0.773	0.541	1.000

Machine Learning

Features

weekly no. of control posts
weekly no. of posts containing ILI related keyword
normalized ILI related post counts
time and seasonal features
- year
- month
- week
- season
lag terms

Model structure

\(Y_w:\) ILI Incidence in week \(w\)

\(X_w:\) Input features obtained in week \(w\)

\[Y_{w+1} = f(X_w, X_{w-1}, X_{w-2})\]

Gradient boosted trees

Sequential learning of weak learners.
Iteratively corrects errors of previous models
Combines predictions using weighted averaging.
Robust to outliers
Handles non-linear relationships

Model evaluation

Time series split validation
- retains temporal information
- mimics continuous data acquisition
Setup:
- initial training: 40 weeks
- test set size: 1 week

Expanding window time series validation ¹

Validation results

Predictions and metrics

Target variable: weeky ILI incidence one week ahead \(Y_{w+1}\)

Metrics

Dataset	MAE*	RMSE
Training	\(23.96\)	\(33.93\)
Validation ⁺	\(56.54\)	\(56.54\)

* Mean absolute error, incidence per 100,000

⁺ mean over all validation runs

Permutation importance of the current model

model agnostic feature importance procedure
random shuffling of single input features

Can “AI” help?

Idea

Filter posts using large a large language model (LLM)

How?

provide case definition in the system prompt
use json structured output option for convenient data processing

Prompt and output

Prompt extract

Analyze the following tweet-like message to determine if it describes the user's own influenza-like illness (ILI). ILI is defined by:

- Fever ≥38°C (100°F) **AND**

- At least one respiratory symptom (cough or sore throat) **PLUS**

- Additional systemic symptoms (headache, muscle aches, chills, fatigue, nasal congestion)

...



{ ... bluesky message dynamically inserted here ... }

// symptom extraction schema
{
    "ili_related" :{
        "type":"bool"
    },
    symptoms:{
        "type":"array",
        "items":{
            "type":"string"
        }
    }
}

// symptom extraction example
{
    "ili_related" :true,
    symptoms:{
      ["fever", "headache"]
    }
}

Extraction using google Gemini API

Examples

ILI positive

Original message

Je suis un peu malade (fièvre, frissons, rhume). Je décide de télétravailler aujourd’hui. Et là, je réalise que les voisins font des travaux. Avec des engins électriques bruyants. Ma tête va exploser !

Machine translation

I'm a little sick (fever, chills, colds). I decide to telework today. And there, I realize that the neighbors do work. With noisy electric vehicles. My head will explode!

LLM symptoms

fièvre,frissons,rhume,tête va exploser

ILI negative

Grippe aviaire : les coupes budgétaires de Trump amplifient la menace pandémique www.lepoint.fr/tiny/1-2586310 #Santé via @lepoint.fr

Aviary flu: Trump's budget cuts amplify the pandemic threat www.lepoint.fr/tiny/1-2586310 #health via @lepoint.fr

nan

LLM annotated post counts

Correlation

	LLM ILI posts	ILI incidence	Control posts
LLM ILI posts	1.000	0.793	0.812
ILI incidence	0.793	1.000	0.557
Control posts	0.812	0.557	1.000

Predictions and metrics

Target variable: weeky ILI incidence \(w_{t+1}\)

Metrics

Dataset	MAE* (LLM filtered)
Training	\(26.64\)
Validation	\(58.87\)

* Mean absolute error, incidence per 100,000

Permutation importance, LLM filtered posts

Conclusion

bluesky = promising data source
more data needed = patience

Outlook

investigate impact of LLM filtering on model performance
modeling of weekly ILI incidence based on message content
continuous data acquisition pipeline (WIP)
User localization based on profile
monitoring of bursts in user activity crucial
repeating the analysis for another country (e.g. Germany)

Pipeline (WIP)

graph LR
    subgraph kestra 
        dlt(dlt) --- posts
        llm --- bqstaging
        llm -- annotation --> bqstaging
        posts --> bqstaging[<b>GBQ</b> \n stage area \n 1 table per kw]
        dlt -- housekeeping --> count
        dlt -- case data --> who_tables
        dlt -- case data --> cdc_tables
        subgraph BigQuery data lake
          bqstaging
          who_tables
          cdc_tables
          count[post counts table]
        end
        bqstaging --- dbt
        who_tables --- dbt
        cdc_tables --- dbt
        count --- dbt
        dbt --> bq[Google \n BigQuery]
        subgraph BigQuery data warehoue
          bq
        end
    end

    bsky[bsky API] --> dlt
    WHO --> dlt
    CDC --> dlt
    bq --> looker[Looker studio \n dashboard]
    bq -- python --> stat1[Statistical analysis]
    bq -- python --> stat2[Machine learning, modeling]

Open source implementation

Data ingestion

python
data load tool (dlt)

Data storage and SQL

Google Big Query

Data modeling

data build tool (dbt)

Workflow orchestration

kestra

available at: https://github.com/kantundpeterpan/bluesky_ddd_influenza

Thank you for your attention.

References

Balduf, Leonhard, Saidu Sokoto, Onur Ascigil, Gareth Tyson, Björn Scheuermann, Maciej Korczyński, Ignacio Castro, and Michaŀ Król. 2024. “Looking at the Blue Skies of Bluesky.” In Proceedings of the 2024 ACM on Internet Measurement Conference, 76–91.

Duarte, Fabio. “Bluesky User Age, Gender, & Demographics (2025).” https://explodingtopics.com/blog/bluesky-users.

“How to Apply Stacking Cross Validation for Time-Series Data? — Datascience.stackexchange.com.” https://datascience.stackexchange.com/questions/41378/how-to-apply-stacking-cross-validation-for-time-series-data.

pythongeeks.org. “Gradient Boosting Algorithm in Machine Learning - Python Geeks — Pythongeeks.org.” https://pythongeeks.org/gradient-boosting-algorithm-in-machine-learning/.

Signorini, Alberto Maria AND Polgreen, Alessio AND Segre. 2011. “The Use of Twitter to Track Levels of Disease Activity and Public Concern in the u.s. During the Influenza a H1N1 Pandemic.” PLOS ONE 6 (5): 1–10. https://doi.org/10.1371/journal.pone.0019467.

Beyond twitter

Outline

Introduction

bluesky: general aspects

Decentralization and Democratization of content algorithms 1

Development of user activity 1

Literature addressing bluesky

Exploration of bluesky data

bluesky API

searchPosts API method

Post metadata

User information

getProfiles API endpoint

Project

Outline

Data extraction

ILI symptom related message extraction

Basal network activity

Case data

Results

Post count time series

Raw posts counts

Keyword posts vs. ILI incidence

Machine Learning

Features

Model structure

Gradient boosted trees

Model evaluation

Validation results

Predictions and metrics

Metrics

Permutation importance of the current model

Can “AI” help?

Idea

How?

Prompt and output

Prompt extract

Examples

ILI positive

Original message

Machine translation

LLM symptoms

ILI negative

LLM annotated post counts

Correlation

Predictions and metrics

Metrics

Permutation importance, LLM filtered posts

Conclusion

Outlook

Pipeline (WIP)

Data ingestion

Data storage and SQL

Data modeling

Workflow orchestration

Thank you for your attention.

References

`bluesky`: general aspects

Decentralization and Democratization of content algorithms ¹

Development of user activity ¹

Literature addressing `bluesky`

`searchPosts` API method

`getProfiles` API endpoint